NVIDIA · alliepiper · Dec 15, 2022 · Jun 29, 2022 · Jun 29, 2022 · Aug 12, 2022
diff --git a/ci/axis/cpu.yml b/ci/axis/cpu.yml
@@ -5,11 +5,9 @@
 
 SDK_TYPE:
   - cuda
-  - nvhpc
 
 SDK_VER:
   - 11.5.1-devel
-  - 22.1-devel-cuda11.5
 
 OS_TYPE:
   - ubuntu
@@ -18,7 +16,6 @@ OS_VER:
   - 20.04
 
 CXX_TYPE:
-  - nvcxx
   - clang
   - gcc
   - icc
@@ -32,55 +29,18 @@ CXX_VER:
   - 10
   - 11
   - 12
-  - 22.1
   - latest
 
 exclude:
-  # Excludes by `SDK_TYPE`.
-  - CXX_TYPE: gcc
-    SDK_TYPE: nvhpc
-  - CXX_TYPE: clang
-    SDK_TYPE: nvhpc
-  - CXX_TYPE: icc
-    SDK_TYPE: nvhpc
-  - CXX_TYPE: nvcxx
-    SDK_TYPE: cuda
-  # Excludes by `SDK_VER`.
-  - SDK_TYPE: cuda
-    SDK_VER: 22.1-devel-cuda11.5
-  - SDK_TYPE: nvhpc
-    SDK_VER: 11.5.1-devel
   # Excludes by `CXX_VER`.
-  - CXX_TYPE: nvcxx
-    CXX_VER: 5
-  - CXX_TYPE: nvcxx
-    CXX_VER: 6
-  - CXX_TYPE: nvcxx
-    CXX_VER: 7
-  - CXX_TYPE: nvcxx
-    CXX_VER: 8
-  - CXX_TYPE: nvcxx
-    CXX_VER: 9
-  - CXX_TYPE: nvcxx
-    CXX_VER: 10
-  - CXX_TYPE: nvcxx
-    CXX_VER: 11
-  - CXX_TYPE: nvcxx
-    CXX_VER: 12
-  - CXX_TYPE: nvcxx
-    CXX_VER: latest
   - CXX_TYPE: gcc
     CXX_VER: 12
-  - CXX_TYPE: gcc
-    CXX_VER: 22.1
   - CXX_TYPE: gcc
     CXX_VER: latest
   - CXX_TYPE: clang
     CXX_VER: 5
   - CXX_TYPE: clang
     CXX_VER: 6
-  - CXX_TYPE: clang
-    CXX_VER: 22.1
   - CXX_TYPE: clang
     CXX_VER: latest
   - CXX_TYPE: icc
@@ -99,6 +59,3 @@ exclude:
     CXX_VER: 11
   - CXX_TYPE: icc
     CXX_VER: 12
-  - CXX_TYPE: icc
-    CXX_VER: 22.1
-
diff --git a/cmake/ThrustCudaConfig.cmake b/cmake/ThrustCudaConfig.cmake
@@ -1,6 +1,6 @@
 enable_language(CUDA)
 
-set(THRUST_KNOWN_COMPUTE_ARCHS 35 37 50 52 53 60 61 62 70 72 75 80 86)
+set(THRUST_KNOWN_COMPUTE_ARCHS 35 37 50 52 53 60 61 62 70 72 75 80 86 90)
 
 # Split CUDA_FLAGS into 3 parts:
 #

diff --git a/dependencies/cub b/dependencies/cub
diff --git a/docs/github_pages/releases.md b/docs/github_pages/releases.md
@@ -8,6 +8,7 @@ nav_order: 3
 
 | Version         | Included In                               |
 |-----------------|-------------------------------------------|
+| 2.0.1           | CUDA Toolkit 12.0                         |
 | 2.0.0           | TBD                                       |
 | 1.17.2          | TBD                                       |
 | 1.17.1          | TBD                                       |

diff --git a/thrust/system/cuda/detail/cdp_dispatch.h b/thrust/system/cuda/detail/cdp_dispatch.h
@@ -42,13 +42,29 @@
  * \endcode
  */
 
-#ifdef THRUST_RDC_ENABLED
+#if defined(CUB_DETAIL_CDPv1)
 
-// seq_impl unused.
+// Special case for NVCC -- need to inform the device path about the kernels
+// that are launched from the host path.
+#if defined(__CUDACC__) && defined(__CUDA_ARCH__)
+
+// seq_impl only used on platforms that do not support device synchronization.
 #define THRUST_CDP_DISPATCH(par_impl, seq_impl)                                \
-  NV_IF_TARGET(NV_ANY_TARGET, par_impl, par_impl)
+  if (false)                                                                   \
+  { /* Without this, the device pass won't compile any kernels. */             \
+    NV_IF_TARGET(NV_ANY_TARGET, par_impl);                                     \
+  }                                                                            \
+  NV_DISPATCH_TARGET(NV_PROVIDES_SM_90, seq_impl, NV_ANY_TARGET, par_impl)
 
-#else // THRUST_RDC_ENABLED
+#else // NVCC device pass
+
+// seq_impl only used on platforms that do not support device synchronization.
+#define THRUST_CDP_DISPATCH(par_impl, seq_impl)                                \
+  NV_DISPATCH_TARGET(NV_PROVIDES_SM_90, seq_impl, NV_ANY_TARGET, par_impl)
+
+#endif // NVCC device pass
+
+#else // CDPv1 unavailable. Always fallback to serial on device:
 
 // Special case for NVCC -- need to inform the device path about the kernels
 // that are launched from the host path.
@@ -62,11 +78,11 @@
   }                                                                            \
   NV_IF_TARGET(NV_IS_HOST, par_impl, seq_impl)
 
-#else // NVCC device pass
+#else // !(NVCC device pass):
 
 #define THRUST_CDP_DISPATCH(par_impl, seq_impl)                                \
   NV_IF_TARGET(NV_IS_HOST, par_impl, seq_impl)
 
 #endif // NVCC device pass
 
-#endif // THRUST_RDC_ENABLED
+#endif // CDP version
+1 −0		.gitignore
+1 −0		CMakeLists.txt
+1 −0		README.md
+1,108 −0		cmake/CPM.cmake
+7 −0		cmake/CubBuildCompilerTargets.cmake
+1 −1		cmake/CubCudaConfig.cmake
+9 −0		cub/detail/detect_cuda_runtime.cuh
+13 −5		cub/detail/device_synchronize.cuh
+34 −13		cub/util_device.cuh
+255 −0		docs/test_overview.md
+125 −18		test/CMakeLists.txt
+163 −0		test/c2h/custom_type.cuh
+384 −0		test/c2h/generators.cu
+102 −0		test/c2h/generators.cuh
+3 −0		test/catch2_runner.cu
+425 −0		test/catch2_test_block_adjacent_difference.cu
+213 −0		test/catch2_test_block_histogram.cu
+326 −0		test/catch2_test_block_load.cu
+520 −0		test/catch2_test_block_merge_sort.cu
+823 −0		test/catch2_test_block_radix_sort.cu
+362 −0		test/catch2_test_block_reduce.cu
+224 −208		test/catch2_test_block_run_length_decode.cu
+536 −0		test/catch2_test_block_scan.cu
+427 −0		test/catch2_test_block_shuffle.cu
+327 −0		test/catch2_test_block_store.cu
+195 −0		test/catch2_test_helper.h
+0 −719		test/test_block_adjacent_difference.cu
+0 −325		test/test_block_histogram.cu
+0 −535		test/test_block_load_store.cu
+0 −443		test/test_block_merge_sort.cu
+0 −641		test/test_block_radix_sort.cu
+0 −767		test/test_block_reduce.cu
+0 −888		test/test_block_scan.cu
+0 −396		test/test_block_shuffle.cu
+23 −258		test/test_util.h
+320 −0		test/test_util_vec.h